Improved Lung Cancer and Mortality Prediction Accuracy Using Survival Models Based on Semi-Automatic CT Image Measurements

Background: In a lung cancer CT screening setting, imaging biomarkers are typically extracted by experienced human readers. We found that adding semi-automatic computer-aided detection (CAD) measurements to a base model significantly improved lung cancer and mortality risk prediction accuracy. Method: Participants' baseline CT scans, characteristics, and 7-year follow-up outcomes were obtained from the National Lung Screening Trial. The selection included all 1810 deceased and a random selection of 4190 surviving participants from the CT screening arm with an image available; the latter subcohort was sampled with replacement up to 24432 to approximate the full CT arm. Seventeen patient characteristics variables endorsed by literature were considered for each model. CAD was used to automatically measure normalized emphysema score, coronary calcium volume, and thoracic aorta calcium volume. Pulmonary nodule consistency, volume, solid core volume (if part-solid), and upper lobe location were annotated by an experienced radiologist with CAD support. Only the largest noncalcified nodule was considered for the model; having no nodules was the reference. Cox proportional hazard regression was performed on patient characteristics variables only (base model) and combined with CAD variables (new model). This was done for three outcomes: lung cancer diagnosis, lung cancer mortality, and overall mortality. The average continuous net reclassification improvements (NRI) between the base and new models were calculated for each year following the baseline scan. To calculate NRI, the net percentages of subjects with and without the event of interest correctly reclassified as high and low risk, respectively, are summed (maximum range: -2 to 2); positive scores indicate that the new model is more accurate. Result: CAD measures were successfully computed for 5575 baseline scans. After sampling, the test cohort consisted of 24370 participants. 3.9% were diagnosed with lung cancer (940/24370) and 6.9% died (1681/24370), of which 24.9% due to lung cancer (418/1681). For all outcomes, the new models were significantly superior to the base model. With lung cancer diagnosis as the outcome, the NRI at 1, 4, and 7 years follow-up were 0.628 (95% confidence interval: 0.373-0.700), 0.331 (0.261-0.390), and 0.349 (0.293-0.389), respectively. The respective NRIs were 0.501 (0.290-0.642), 0.288 (0.221-0.374), and 0.255 (0.218-0.339) when predicting lung cancer mortality and 0.496 (0.295-0.610), 0.301 (0.239-0.376), and 0.270 (0.201-0.320) when predicting overall mortality. Conclusion: CAD measures of emphysema and atherosclerosis and CAD-supported pulmonary nodule annotations are of added value for predicting lung cancer and mortality. These new models may be used to further personalize lung cancer CT screening follow-up protocols.